It’s Technically Fair, But You Might Not Like It

• In Event: Algorithms at Work: The Practice of Prediction - III: Reframing Algorithmic Accountability in Practice

Abstract

Filtering, curation, and decision-making about seemingly everything is being taken over by digital computing. In this context, "fair algorithms" has become a technical problem, a legal problem, and a rallying cry. Machine learning research now quantifies fairness as constraints and metrics within and about the algorithmic systems that select job applicants, predict recidivism, offer housing, find rides, and filter social media. Yet the fairnesses addressed so far in what is known to insiders as the "FAT*" (fairness, accountability, transparency) domain have mostly been arid (Mitchell, Potash, & Barocas 2018). "Fair" means either compliance with particular US laws about housing and civil rights, or it means statistical fairness—a mathematical property of equality. This paper argues that computer systems abiding by these fairnesses will still be widely perceived as unfair. The problem: an over-reliance on a Ralwsian conception of justice. As an alternative, the paper argues that algorithm designers need to entangle computing with the morass of applied ethics and lived experience (after Suchman 2006). To that end, the paper proposes a typology of the kinds of fairness found in applied ethics and STS that are relevant for the people who operate algorithmic platforms that curate, filter, or predict. Examples are drawn from the literature about non-computer technologies to show that these "neglected fairnesses" are actually quite normal in engineering, and are sometimes framed as technical, often via the idea of "safety." In contrast, within computing these are currently framed as social, and therefore outside the scope of serious consideration.

Author

• Christian Sandvig, University of Michigan